Investigate ECG Signal Classifications Using Grad-CAM

This example shows how to use the gradient-weighted class activation mapping (Grad-CAM) technique to understand the classification decisions of a deep neural network for ECG signal classification. This example is step eight in a series of examples that take you through an ECG signal classification workflow. This example follows the Uncertainty Quantification for ECG Signal Classification example. For more information about the full workflow, see ECG Signal Classification Using Deep Learning.

To run this example, open ECG Signal Classification Using Deep Learning and navigate to scripts\S8_ExplainPredictions. Alternatively, if you already have MATLAB open, then run

openExample("deeplearning_shared/ECGSignalClassificationUsingDeepLearningExample")

This project contains all of the steps for this workflow. You can run the scripts in order or run each one independently.

Grad-CAM uses the gradient of the classification score with respect to the convolutional features determined by the network to understand which parts of the data are most important for classification. For time-series data, Grad-CAM computes the most important time steps for the classification decision of the network. For example, this image shows an example of an atrial fibrillation signal. The color of the signal represents the parts of the signal that the Grad-CAM explanation determines the network uses for classification.

The ECG signal features highlighted by Grad-CAM represent the specific patterns or intervals that the neural network relies upon to make its prediction. However, it is important to note that these features are not necessarily the same as those a physician would use when diagnosing atrial fibrillation. While a clinician might focus on established criteria such as irregular R-R intervals or the absence of P-waves, the network might identify subtle, data-driven cues that are not part of traditional medical knowledge.

In this example, you will investigate ECG signal classifications using Grad-CAM.

Load the test data. If you have run the previous step, Prepare Data for ECG Signal Classification, then the example uses the data that you prepared in that step. Otherwise, the example prepares the data as shown in Prepare Data for ECG Signal Classification.

if ~exist("XTest","var") || ~exist("TTest","var")
    [~,~,~,~,XTest,TTest] = prepareECGData;
end

Load a pretrained network. If you have run the previous training step, Improve Adversarial Robustness of Deep Learning Network for ECG Signal Classification, then the example uses your trained network. Otherwise, load a pretrained network. The network has been trained using the steps shown in Improve Adversarial Robustness of Deep Learning Network for ECG Signal Classification.

if ~exist("netRobust","var")
    load("adversariallyTrainedECGNetwork.mat");
end

numToPlot = 2;

scores = minibatchpredict(netRobust,XTest,InputDataFormats="CTB");
YTest = scores2label(scores,categories(TTest));

isTruePositive = (YTest == "A") & (TTest == "A");
idxTruePositive = find(isTruePositive,numToPlot);

figure
tiledlayout("flow")

for ii = 1:numToPlot
    currentSignal = XTest{idxTruePositive(ii)};
    currentLabel = YTest(idxTruePositive(ii));

    importance = gradCAM(netRobust, currentSignal, currentLabel);

nexttile
    plotWithColorGradient(currentSignal(1:500), importance(1:500))
end

isFalseNegative = (YTest ~= "A") & (TTest == "A");
idxFalseNegative = find(isFalseNegative,numToPlot);

figure
tiledlayout("flow")

for ii = 1:numToPlot
    currentSignal = XTest{idxFalseNegative(ii)};
    currentLabel = YTest(idxFalseNegative(ii));

    importance = gradCAM(netRobust,currentSignal,currentLabel);

nexttile
    plotWithColorGradient(currentSignal(1:500),importance(1:500))
end

function plotWithColorGradient(sequence,importance)

x = 1:size(sequence,2) + 1;
y = [sequence NaN];
c = [importance NaN];

patch(x,y,c,EdgeColor="interp");

cb = colorbar(gca);
cb.Label.String = "Importance";
cb.Label.Rotation = 270;

xlabel("Samples")
ylabel("Amplitude (mV)")

box on
xlim tight
ylim padded

end

See Also

gradCAM | imageLIME

Topics

• Battery State of Charge Estimation Using Deep Learning
• Grad-CAM Reveals the Why Behind Deep Learning Decisions
• Interpret Deep Learning Time-Series Classifications Using Grad-CAM
• Explore Semantic Segmentation Network Using Grad-CAM